Bias Control and Injection for Multi-Agent LLM Aggregation
October 16, 2026
A new test-time aggregation framework manages multi-agent disagreement by using style-normalized reasoning traces for bias control and intentional bias injection to signal decision instability. This method improves answer selection without requiring validation data or prior knowledge of agent strength.
HOW THIS AFFECTS YOU
●
builderYou can implement this framework to improve the reliability of multi-agent reasoning systems in production.
●
researcherYou can use bias injection as a reliability signal to quantify agent decision instability.